Skip to content

Latest commit

 

History

History
94 lines (72 loc) · 14.3 KB

MODEL_ZOO.md

File metadata and controls

94 lines (72 loc) · 14.3 KB
Raw

Model Zoo

For the leaderboard on public benchmarks, please refer to LEADERBOARD.md.

Note: all the models below are selected by the performance on validation sets.

Unsupervised learning (USL) on object re-ID

  • Direct infer models are directly tested on the re-ID datasets with ImageNet pre-trained weights.

Market-1501

Method Backbone Pre-trained mAP(%) top-1(%) top-5(%) top-10(%) Train time Download
Direct infer ResNet50 ImageNet 2.2 6.7 14.9 20.1 n/a
UDA_TP ResNet50 ImageNet 34.7 58.6 74.0 78.9 ~2h [config] [model]
strong_baseline ResNet50 ImageNet 70.5 87.9 95.7 97.1 ~2.5h [config] [model]
MMT ResNet50 ImageNet 74.3 88.1 96.0 97.5 ~4.5h [config] [model]
SpCL ResNet50 ImageNet 76.0 89.5 96.2 97.5 ~2h [config] [model]

DukeMTMC-reID

Method Backbone Pre-trained mAP(%) top-1(%) top-5(%) top-10(%) Train time Download
Direct infer ResNet50 ImageNet 2.3 7.5 14.7 18.1 n/a
UDA_TP ResNet50 ImageNet 42.3 64.4 76.0 79.9 ~2h [config] [model]
strong_baseline ResNet50 ImageNet 54.7 72.9 83.5 87.2 ~2.5h [config] [model]
MMT ResNet50 ImageNet 60.3 75.6 86.0 89.2 ~4.5h [config] [model]
SpCL ResNet50 ImageNet 67.1 82.4 90.8 93.0 ~2h [config] [model]

... (TBD)

Unsupervised domain adaptation (UDA) on object re-ID

  • Direct infer models are trained on the source-domain datasets (source_pretrain) and directly tested on the target-domain datasets.
  • UDA methods (MMT, SpCL, etc.) starting from ImageNet means that they are trained end-to-end in only one stage without source-domain pre-training.

DukeMTMC-reID -> Market-1501

Method Backbone Pre-trained mAP(%) top-1(%) top-5(%) top-10(%) Train time Download
Direct infer ResNet50 DukeMTMC 27.2 58.9 75.7 81.4 ~1h [config] [model]
UDA_TP ResNet50 DukeMTMC 52.3 76.0 87.8 91.9 ~2h [config] [model]
strong_baseline ResNet50 ImageNet 75.6 90.9 96.6 97.8 ~3h [config] [model]
MMT ResNet50 ImageNet 80.9 92.2 97.6 98.4 ~6h [config] [model]
SpCL ResNet50 ImageNet 78.2 90.5 96.6 97.8 ~3h [config] [model]

Market-1501 -> DukeMTMC-reID

Method Backbone Pre-trained mAP(%) top-1(%) top-5(%) top-10(%) Train time Download
Direct infer ResNet50 Market 28.1 49.3 64.3 69.7 ~1h [config] [model]
UDA_TP ResNet50 Market 45.7 65.5 78.0 81.7 ~2h [config] [model]
strong_baseline ResNet50 ImageNet 60.4 75.9 86.2 89.8 ~3h [config] [model]
MMT ResNet50 ImageNet 67.7 80.3 89.9 92.9 ~6h [config] [model]
SpCL ResNet50 ImageNet 70.4 83.8 91.2 93.4 ~3h [config] [model]

... (TBD)